Multiple Kernel Learning Based One-class Support Vector Machine

The traditional one-class support vector machine(OCSVM)has obtained better performance in the field of unsupervised learning and thus gained more and more attentions.However,there exist several shortcomings for OCSVM.First,the classification performance of OCSVM largely depends on the choice of kernel function and its parameters.If they are inappropriately chosen,the classification result of OCSVM will be poor.Unfortunately,there exists no reliable theoretical reference on choosing kernel function and its parameters for OCSVM up to now.Second,OCSVM is very sensitive to the noise among the given training set.If the number of noise in training set is larger,the performance of OCSVM will be seriously affected.Third,towards combined data,e.g.,the patient information in medical diagnosis containing not only the discrete or continuous numerical information of blood type,blood pressure and etc.,but also the image information on CT,color doppler ultrasonic and etc.,OCSVM cannot deal with them at all.To solve the abovementioned problems,multiple kernel learning(MKL)can be introduced into OCSVM.MKL has obtained better performance in the classification and regression tasks than the single kernel methods.To deal with the aforementioned problems for OCSVM and simultaneously enhance its classification performance,two novel methods of multiple kernel OCSVM are proposed.1.The centered kernel alignment based multiple kernel OCSVM is proposed.First,centered kernel alignment(CKA)is utilized to calculate the weights for every kernel matrix.The obtained weights are then used as the linear combination coefficients.Furthermore,the kernel functions in different types are linearly combined to construct the combined kernel function.Finally,the combined kernel function is introduced into the dual optimization problem of OCSVM to substitute the single kernel function.The proposed method can not only avoid the problem of choosing kernel function,but also improve the generalization performance and anti-noise ability.The experiments are conducted on the benchmark data sets to compare the proposed method and its five related approaches.The efficiency of the proposed method is thus validated.2.The deep multiple kernel OCSVM is proposed.Deep multiple kernel learning(DMKL)is utilized to combine the given kernel functions.That is to say,the combined kernel matrix is obtained via multiple feature mapping on several kernel functions.Then,the gradient descent method is used to achieve the optimal combination weights.Furthermore,the optimal combined kernel function can be obtained.Finally,the combined kernel function is also introduced into the dual optimization problem of OCSVM to substitute the single kernel function.The proposed method can not only avoid the problem of choosing kernel function and its parameters,but also improve the generalization ability of the model.It can be found from the experimental results on the benchmark data sets that the proposed method obtains better classification performance in comparison with its five related approaches.